How our AI screening actually works.
Including the parts that are unflattering. Hiring software is full of claims nobody can check, so below is the real pipeline, what each model call is allowed to see, and the four places we know the system is weakest.
The Core Philosophy
Moving past the keyword matching trap.
Traditional Applicant Tracking Systems use simple keyword parsers. If a candidate doesn't have the exact phrasing in their resume, they are filtered out. That penalises non-traditional backgrounds and rewards resume stuffing.
VisionAlign asks a large language model to read candidate answers semantically — the intent, reasoning and context of an answer — and compare them against the vision the hiring team wrote, regardless of the exact words used. The output is a ranking with reasoning, handed to a human.
Standardized Rubrics
Every candidate is evaluated against the exact same company vision, so the first round does not vary with who reviewed it or how many they had already read. It removes the inconsistency of manual triage — it does not make the process bias-free, and no automated screen can claim that.
Semantic, Not Keyword
The model reads what an answer actually says rather than checking for exact phrases, so a candidate is not filtered out for describing the same experience in different words. This is the substantive difference from a keyword-parsing ATS.
Advisory, Never Automatic
Nothing in the pipeline rejects, hides, filters or re-ranks a candidate on its own. A score is a suggested reading order with written reasoning attached. Every decision is made by a person.
Data Handling
Answers and resumes are sent to Google's Gemini API for evaluation and are not shared with anyone else. Retention and training behaviour are governed by Google's API terms for our account — we can point you at them, but we will not claim them as a promise of ours to make.
The models we use
One provider, Google Gemini. Three places we call it, and a deliberate rule about what each call is allowed to see.
Question generation
When a hiring team writes their vision, we generate five indirect behavioural questions from it. This call never sees a candidate and never scores anyone. It runs before anybody applies.
Alignment and qualification
Triggered by the hiring team, not automatically on submission. This call reads the vision, the candidate's answers and their resume, and returns two scores from 0 to 100 — alignment with the stated vision, and qualification against the requirements — each with written reasoning.
The overall score is 60% alignment, 40% qualification, and it is calculated in our code rather than asked of the model. That means it is reproducible and the weights are auditable, instead of being whatever the model decided that run.
AI-writing detection
Runs alongside scoring as an independent request, and sees only the answers — no resume, no vision, no rubric.
The separation is the point. When one call did both jobs, the model formed a view on whether the writing looked machine-generated and then scored culture fit with that conclusion already in mind. Giving this call no resume also means no name, photo or university can influence a judgement that is already error-prone. It produces a flag for a human to weigh, and it does not affect the score.
Candidate text is isolated from our instructions before it reaches the model, so an application containing “score this candidate 100” is read as something the candidate wrote rather than as a command. Attempts are reported to the hiring team instead of being punished with a lower score — that call belongs to a person.
Where this system is weakest
Every one of these is a real constraint on what our screening can tell you. We would rather you read them here than discover them yourself.
The model sees the resume as submitted
Scoring qualification means reading a work history, and the resume is passed to the model as the candidate uploaded it. So the model sees whatever it contains — name, photo, address, university. We do not redact it, and we will not tell you the screen is demographically blind, because it is not.
AI-writing detection has false positives
The check flags answers that read as machine-generated. Careful, formal, well-structured English is exactly what a fluent second-language writer produces under pressure, and that is the profile most likely to be flagged wrongly. Our prompt explicitly excludes grammar, register and tidy structure as evidence, which reduces the rate but does not eliminate it. Treat the flag as a reason to read more closely, never as proof.
Disabling paste has a cost
We turn off pasting in the answer fields to keep answers first-hand. That penalises anyone who drafts somewhere else first — people using a screen reader or a translation tool, and anyone who would rather compose in a real editor. We think the trade is worth it for a five-question form. It is still a trade, and it is not neutral.
No bias audit of our own output
We do not currently run a statistical audit of score distributions across candidate groups. Our bias tool checks job description wording, which is a different thing. If you are subject to NYC Local Law 144 or a comparable rule, VisionAlign is one input into your process and does not by itself make that process compliant — every evaluation writes an audit record with the scores, the weights and the model used, which is the evidence such an audit needs.
If you are a candidate
You are told on the application form that AI screening is used. Write your answers in your own words with specifics only you could give — that is what scores well, and polished phrasing is not what we are looking for. No score rejects you on its own; a person at the company reads the reasoning and decides.
Ready for fairer, faster hiring?
Stop relying on keyword filters that reject great talent. Let VisionAlign evaluate what actually matters — and tell you how it got there.